基于改进的YOLOv5的航拍图像中小目标检测算法*杨慧剑,孟亮(太原理工大学信息与计算机学院,山西晋中030600)摘要:目前基于无人机航拍的目标检测技术广泛应用于军事和民用领域,但因其存在成像距离远、高空拍摄图像模糊和目标信息占比小等问题,目标检测准确率不高。针对这一问题,提出一种基于YOLOv5的改进算法。该算法首先在数据增强方面对原始图像进行加雾处理,提高其在雾天的鲁棒性;其次通过融合CBAM模块,来增强不同通道和空间的重要性;再者将原算法中的SPP更换为ASPP,以减小池化操作对特征信息的影响;最后在FPN结构中增加一层检测头,用于更细粒度的检测目标。以YOLOv5s为Baseline,实验表明,改进后的算法比原算法的mAP_0.5提高了6.9%,可以有效应用于航拍小目标的检测。关键词:YOLOv5;无人机;注意力机制;金字塔池化;特征金字塔中图分类号:TP391.4文献标志码:Adoi:10.3969/j.issn.1007-130X.2023.06.013AsmalltargetdetectionalgorithmbasedonimprovedYOLOv5inaerialimageYANGHui-jian,MENGLiang(SchoolofInformationandComputer,TaiyuanUniversityofTechnology,Jinzhong030600,China)Abstract:Atpresent,thetargetdetectiontechnologybasedonUAVaerialphotographyiswidelyusedinmilitaryandcivilfields,buttheaccuracyoftargetdetectionisnothighbecauseofthelongimag-ingdistance,blurredimagestakenathighaltitudes,andsmallproportionoftargetinformation.Tosolvethisproblem,animprovedalgorithmbasedonYOLOv5isproposed.Firstly,theoriginalimageisfoggedtoimproveitsrobustnessonfoggydays.Secondly,theimportanceofdifferentchannelsandspacesisenhancedthroughtheintegrationofCBAMmodules.Furthermore,theSPPintheoriginalalgorithmisreplacedbytheASPPtoreducetheinfluenceofpoolingoperationonfeatureinformation.Finally,adetectionheadisaddedtotheFPNstructuretodetecttargetswithfinergranularity.TakingYOLOv5sasbaseline,theexperimentprovesthattheimprovedalgorithmincreasesmAP_0.5by6.9%incomparisontotheoriginalalgorithm,andcanbeeffectivelyappliedtothedetectionofsmalltargetsinaerialphotography.Keywords:YOLOv5;unmannedaerialvehicle(UAV);attentionmechanism;spatialpyramidpool-ing;featurepyramid1引言近几年无人驾驶飞机在低空、超低空等领域得到长足的发展,应用范围也逐步扩大,机器也由高精尖向平民化迈进。随着...
